[kroaton]

Which would be very interesting to test, as larger models (such as Deepseek V4 Flash or Qwen 397B) seem to compress better. Their Q2 quants are usable as is, even without the ternary compression.